Wang Yeao, Rong Jianke, Wei Zhili, Bai Xiaoyu, Deng YunDan
The First Clinical Medical College, Lanzhou University, Lanzhou, Gansu, China.
The Cardiovascular Surgery, The First Hospital of Lanzhou University, Lanzhou, Gansu, China.
Front Cardiovasc Med. 2025 Jun 25;12:1590367. doi: 10.3389/fcvm.2025.1590367. eCollection 2025.
Currently, heart failure has become one of the major complications in the advanced stages of various cardiovascular diseases. Numerous predictive models have been developed to estimate the mortality rate of heart failure patients; however, these models often require the measurement of multiple indicators and the inclusion of various scoring systems. Critically ill patients are often unsuitable for extensive diagnostic tests, and many primary care hospitals lack comprehensive diagnostic equipment. In contrast, blood tests are not only simpler but also reflect the overall health status of the body. Therefore, using simpler methods to predict mortality in intensive care unit (ICU) patients has become the focus of this study.
A total of 5,383 cases from the eICU database were utilized for model development, while 530 cases from the MIMIC-IV database were employed for external testing. The patients were primarily diagnosed with heart failure, and the data included demographic information, blood oxygen saturation, white blood cells, red blood cells, platelets, hemoglobin, electrolytes, lactate, glucose, and other biochemical and physiological indicators collected during the ICU stay. Enhance the accuracy of data analysis and improve the universality of the model, all data underwent rigorous preprocessing prior to training, combined with data standardization. We utilized a variety of machine learning algorithms for modeling purposes, including Logistic Regression (LR), Support Vector Machine (SVM), Decision Trees, Random Forests, Gradient Boosting Machines (GBM), XGBoost, and Neural Networks. The performance of the model was assessed through cross-validation and evaluated using the F1-score.
Through feature selection, 15 key variables were ultimately identified. Among the nine machine learning models evaluated, the Multilayer Perceptron (MLP) demonstrated the best overall performance. In predicting mortality (i.e., the deceased population), the MLP achieved an F1 score of 0.54, a recall of 0.71, and a precision of 0.44. The relatively high F1 score of the MLP highlights its potential clinical application value.
目前，心力衰竭已成为各种心血管疾病晚期的主要并发症之一。已经开发了许多预测模型来估计心力衰竭患者的死亡率；然而，这些模型通常需要测量多个指标并纳入各种评分系统。危重症患者往往不适合进行广泛的诊断测试，而且许多基层医院缺乏全面的诊断设备。相比之下，血液检测不仅更简单，还能反映身体的整体健康状况。因此，使用更简单的方法预测重症监护病房（ICU）患者的死亡率已成为本研究的重点。
共使用了eICU数据库中的5383例病例进行模型开发，同时使用MIMIC-IV数据库中的530例病例进行外部测试。患者主要诊断为心力衰竭，数据包括人口统计学信息、血氧饱和度、白细胞、红细胞、血小板、血红蛋白、电解质、乳酸、葡萄糖以及在ICU住院期间收集的其他生化和生理指标。为提高数据分析的准确性并提高模型的通用性，所有数据在训练前都经过了严格的预处理，并结合了数据标准化。我们使用了多种机器学习算法进行建模，包括逻辑回归（LR）、支持向量机（SVM）、决策树、随机森林、梯度提升机（GBM）、XGBoost和神经网络。通过交叉验证评估模型的性能，并使用F1分数进行评价。
通过特征选择，最终确定了15个关键变量。在评估的9种机器学习模型中，多层感知器（MLP）表现出最佳的整体性能。在预测死亡率（即死亡人群）方面，MLP的F1分数为0.54，召回率为0.71，精确率为0.44。MLP相对较高的F1分数突出了其潜在的临床应用价值。
